nvcc compilation time

Accelerated Computing CUDA CUDA Programming and Performance
1

Few questions on nvcc compilation time.

(1) Every time when I compile my code, nvcc generates warning messages for four times. This is quite strange. I don’t know the reason for this.

The compiler command that I used is the following

/usr/local/cuda/bin/nvcc  -gencode=arch=compute_10,code=\"sm_10,compute_10\"  -gencode=arch=compute_20,code=\"sm_20,compute_20\"  -m64 --compiler-options -fno-strict-aliasing  -I. -I/usr/local/cuda/include -I../../common/inc -I../../../shared//inc -DUNIX -O2   -o obj/x86_64/release/units.cu.o -c units.cu

Any reason on why I get the same warning messages for four times ?

(2) I also see that the compilation time is quite high when compared with pure c++/g++ compilation.

For example, if I take a test.cpp file and rename it as test.cu file, and compile it with nvcc the compilation time is about 5 times more.

Is there anyway to reduce the compilation time using nvcc ?

Thanks.

2

[font=“Courier New”]-gencod[/font]e results in a separate compiler invocation for each of the given [font=“Courier New”]code=[/font] arguments. So in your case the compiler is invoked four times to compiles the device code and a fifth time to generate the host code. If a warning is provoked by device code, it is thus emitted four times.

You can remove architectures you are not using to speed up compilation.

3

That is right. Thanks for the clarification. I did not realize that the default Makefile that comes with the SDK examples generates code in sm_10, sm_20, compute_10, compute_20.

I turned off the compilation for all other modes except sm_20.

The overall compilation speed issue still remains (nvcc vs g++). But not too critical at this point.